One hundred and eleven of these 112 sites send the exact same Strict-Transport-Security header: max-age=31536000. Not one of them sends includeSubDomains. Not one sends preload. And not one sends any other security header at all — no Content-Security-Policy, no X-Content-Type-Options, no Referrer-Policy, no X-Frame-Options. The whole defensive posture of the fleet is one line, copied 111 times.
That would be a boring inventory post. What makes it worth writing is the second number: 28 of those 112 domains answer a plain http:// request with 200 OK and a full page — and I hashed them, and all 28 bodies are byte-for-byte identical to what the same URL returns over HTTPS. A quarter of the network is still serving its content over cleartext, while telling the browser, on that same cleartext response, to please always use HTTPS.
The browser is required to ignore that instruction. That is the interesting bug.
What I actually measured
The corpus is 114 market and forecast sites I operate — coin pages, perp funding tables, prediction-market write-ups, plus the methodology notes that explain where a forecast number comes from. All of them sit behind the same CDN, all are static HTML, and all were probed on 2026-11-18 from one machine in the same half hour.
Two of the 114 did not answer at all. exchangeguidehub.com and tokenfundingindex.com returned SSLEOFError on every attempt, and both are NXDOMAIN at the resolver — they no longer exist in DNS. I excluded them, which leaves 112.
For each of the 112 I did four requests:
-
GET https://domain/— record the response headers. -
GET http://domain/— plaintext, redirects not followed, so I can see whether the edge upgrades or answers. -
GET http://www.domain/andGET https://www.domain/— same question for thewwwhost, which is a separate hostname with its own configuration.
No browser, no Lighthouse, no header-grading SaaS. Just the wire, because every claim below is a claim about what the server puts on the wire.
The entire security posture is one header
I looked for 12 headers. Here is the full result across 112 live hosts:
| Header | Sites sending it |
|---|---|
Strict-Transport-Security |
111 |
Content-Security-Policy |
0 |
Content-Security-Policy-Report-Only |
0 |
X-Frame-Options |
0 |
X-Content-Type-Options |
0 |
Referrer-Policy |
0 |
Permissions-Policy |
0 |
Cross-Origin-Opener-Policy |
0 |
Cross-Origin-Embedder-Policy |
0 |
Cross-Origin-Resource-Policy |
0 |
Access-Control-Allow-Origin |
0 |
X-XSS-Protection |
0 |
Eleven headers at zero. The twelfth at 111 of 112, with a single value: max-age=31536000. The one holdout sends nothing at all — even its privacy page comes back as 2,680 bytes of HTML with no transport directive on it.
This is what a copied template looks like from the outside. Someone configured a Cloudflare rule once, it produced exactly one directive, and the directive was replicated across a fleet without anyone asking what else the fleet should be sending. The absence is uniform, which is much stronger evidence of a shared default than of 112 individual decisions.
For dev.to readers the X-Content-Type-Options: nosniff line is the one that stings. Without it, a browser is allowed to sniff a response's type instead of trusting Content-Type. On a static site that serves only .html, .css and .js this is mostly theoretical — but "mostly theoretical" is the same category as the rest of this table, and nosniff costs one line in a config file.
The preload gap is two keywords wide
max-age=31536000 is not a random number. It is exactly 365 days, which is the minimum Chromium requires for HSTS preload eligibility. Whoever wrote the config read the requirement list and got the hard part right.
The Chromium preload list asks for four things:
- A valid certificate.
- A
301redirect fromhttp://tohttps://on the same host. -
includeSubDomains. -
preload, withmax-ageof at least 31,536,000.
The fleet meets (1) and, on 84 of 112 hosts, (2). It meets (3) on zero hosts. It meets (4) on zero hosts. So the entire fleet sits two keywords away from a browser-baked-in guarantee, and those two keywords were never added.
The practical difference is who does the remembering. With preload, the browser knows before the first request and never sends cleartext. Without it, the first request to a host is cleartext by definition, and the HSTS header it receives is what teaches the browser to upgrade from then on. Every protection in this scheme starts on visit two.
Which brings us to the part that is not theoretical.
28 sites return the same page over http://
Of the 112 live hosts, 84 answer http://domain/ with a 301 to HTTPS. That is correct behavior, and it is what I assumed the whole fleet did.
The other 28 answer 200 OK. Not a stub, not an interstitial, not a redirect page — the site. I fetched each one twice, once over plaintext and once over TLS, and hashed both bodies:
import hashlib, requests
UA = {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) "
"AppleWebKit/537.36 (KHTML, like Gecko) "
"Chrome/141.0.0.0 Safari/537.36"}
def compare(domain):
plain = requests.get("http://%s/" % domain, headers=UA,
timeout=20, allow_redirects=False)
tls = requests.get("https://%s/" % domain, headers=UA, timeout=20)
return {
"plain_status": plain.status_code,
"plain_bytes": len(plain.content),
"tls_bytes": len(tls.content),
"same_body": (hashlib.sha256(plain.content).hexdigest()
== hashlib.sha256(tls.content).hexdigest()),
"plain_hsts": plain.headers.get("Strict-Transport-Security"),
}
Twenty-eight out of twenty-eight came back same_body: True. Same SHA-256, same byte count:
| Host | Cleartext bytes | Identical to HTTPS |
|---|---|---|
| coinatlaslab.com | 220,999 | yes |
| perpfundingwatch.com | 145,725 | yes |
| coinwikiworld.com | 137,705 | yes |
| coinwikicenter.com | 128,063 | yes |
| web3academyhub.com | 116,701 | yes |
| chaintickerzone.com | 96,111 | yes |
| perptraderjournal.com | 98,143 | yes |
| polymarkettutorials.com | 18,246 | yes |
That is 2,534,288 bytes of production HTML available with no TLS at all, median 95,921 bytes per page. It includes small pages too — a 14 KB beginner's guide to Arbitrum vs Base is served identically over both schemes, because the omission is per-hostname, not per-page-size.
The header is sitting on the response where it gets ignored
Here is the part I did not expect. On 27 of those 28 cleartext 200 responses, the server also sends Strict-Transport-Security: max-age=31536000.
RFC 6797 §8.1 is blunt about this: a user agent must ignore the STS header when it arrives over a non-secure transport. The reason is obvious once you say it out loud — if an on-path attacker can rewrite the response, they can also delete the header, so a cleartext response can never be trusted to establish transport policy.
So the sequence for a first-time visitor to one of these 28 hosts is:
1. GET http://coinatlaslab.com/ -> 200 OK, 220,999 bytes of HTML
2. Response header: Strict-Transport-Security: max-age=31536000
3. Browser: ignored (RFC 6797 8.1, non-secure transport)
4. Visitor reads the page over cleartext. HSTS was never established.
Every subsequent visit is only protected if the visitor happens to type https:// or the site's internal links are absolute https://. The visitors who never do that are, functionally, on an unencrypted site full time.
And the 84 hosts that do redirect are not much better off in the first round trip: a 301 from http:// is itself a cleartext response an attacker can replace with their own page. The redirect closes the window for repeat visitors, not for the first request. That is precisely why preload exists, and precisely why it is missing here.
The split follows the deployment batch, not the intent
The 28 non-upgrading hosts are not scattered. Grouped by domain family:
| Family | Answers http:// with 200 |
Upgrades with 301 |
|---|---|---|
| coinatlas* | 4 | 0 |
| perp* | 5 | 0 |
| altcoin* | 3 | 0 |
| coinwiki* | 3 | 2 |
| token* | 3 | 0 |
| chain* | 2 | 0 |
| dex* | 2 | 0 |
| spacex* | 0 | 15 |
| openai* | 0 | 8 |
| prediction* | 0 | 8 |
| event* | 0 | 6 |
| market* | 1 | 24 |
spacex* is 15 for 15 in upgrading. perp* is 5 for 5 in not upgrading. coinatlas* is 4 for 4. That is not a per-site misconfiguration pattern; that is two deployment batches with two different templates, and one of the templates never got the "Always Use HTTPS" setting.
I can't prove the batch story from the outside — I'm inferring it from the naming, which is the weakest kind of evidence in this post. But the clustering is too clean for chance: perp and coinatlas and altcoin and token and chain and dex are all 100% non-upgrading, while spacex and openai and prediction and event are all 100% upgrading. Something in how those two groups were provisioned differs, and it isn't the content.
www is a second host with its own config
The www hostname is where a copied config usually diverges from its parent, so I probed it separately.
85 of 112 www hosts don't exist at all — plaintext returns 503, and TLS doesn't connect. Of the 27 that do exist, 24 return 404 over HTTPS and 3 return 200. Twenty-six of those 27 send their own Strict-Transport-Security header, which means subdomain coverage here is not inherited from the apex; it's re-declared per hostname.
But zero of the 27 upgrade a plaintext request. http://www.domain/ returns 404 or 200 directly, never a 301. So the www hosts, which do carry the header, are in exactly the same state as the 28 apex hosts: policy declared on the cleartext channel, redirect absent.
Worth noting what this implies for includeSubDomains. If the apex directive had it, none of this per-host repetition would be needed for policy — the browser would refuse cleartext for every subdomain, before the request leaves the machine. Because it doesn't, coverage depends on 27 separate hostname configurations each staying correct forever. One of them, dexnewshub.com, already isn't sending anything.
Nothing stateful to leak, which is the one thing that went right
While I had the headers open I counted Set-Cookie on the same 112 responses.
Zero. Not one site sets a cookie on the homepage. Not a session id, not a consent flag, not a cf_* value. All 112 responses are also cf-cache-status: DYNAMIC with Vary: Accept-Encoding, so nothing is being cached at the edge in a stateful way either.
This is good, and it is probably accidental. No cookies means no cookie-scope mistakes, no SameSite gaps, no session fixation surface, and no cache-poisoning class of bug where a personalized response gets stored. But it also means the only per-visitor state in the whole fleet lives in localStorage and in the third-party script these pages load, which is a different audit for a different day. The header layer being clean of state does not mean the site is clean of state.
What I'd change, in order
If you are auditing a fleet and the table above looks like yours, the ordering matters more than the checklist:
1. Turn on the HTTP→HTTPS redirect before touching anything else. It is one setting, it has no compatibility risk, and on this fleet it is 28 hosts wide. Until it exists, HSTS is decoration.
2. Add includeSubDomains; preload to the existing header. The max-age is already at the required floor. Two keywords, and the protection stops depending on a browser having visited before. Check the preload requirement list before submitting — it includes the redirect from step 1, which is why step 1 goes first.
3. Add X-Content-Type-Options: nosniff. One line, no policy to tune, no way to break a page that serves correct content types.
4. Add frame-ancestors 'none' via CSP, not X-Frame-Options. Same clickjacking protection, and it doesn't need the legacy header alongside it.
5. Only then attempt a real CSP, in report-only first. This is the step that actually takes time, because it needs a report collector and a week of real traffic. default-src 'self' deployed straight to production on a site with inline scripts and a third-party tracker is how you get a blank page and a bad afternoon.
A verification pass that doesn't need a browser, and that you can run in CI against a list of hosts:
import requests
REQUIRED = {
"Strict-Transport-Security": ("max-age=", "includeSubDomains", "preload"),
"X-Content-Type-Options": ("nosniff",),
"Referrer-Policy": (),
"Content-Security-Policy": ("frame-ancestors",),
}
def audit(domain):
problems = []
plain = requests.get("http://%s/" % domain, timeout=20,
allow_redirects=False, headers={"User-Agent": "audit/1.0"})
if plain.status_code == 200:
problems.append("cleartext 200: no HTTP->HTTPS redirect")
elif plain.status_code not in (301, 308):
problems.append("cleartext %d: unexpected" % plain.status_code)
tls = requests.get("https://%s/" % domain, timeout=20,
headers={"User-Agent": "audit/1.0"})
for name, tokens in REQUIRED.items():
value = tls.headers.get(name)
if value is None:
problems.append("missing %s" % name)
continue
for token in tokens:
if token.lower() not in value.lower():
problems.append("%s lacks %r" % (name, token))
return problems
Run it against the 112 hosts and it returns 28 lines starting with cleartext 200, 112 lines about the CSP, and 112 about preload. The list is the work.
What this sample can't tell you
Three limits worth stating, because the numbers above are easy to over-read.
I measured the header layer only. Whether a browser actually enforces a policy, whether a nosniff would have blocked a specific resource, whether any of these pages would break under frame-ancestors — none of that is visible from requests. Header presence is a necessary condition, not a working defense.
It is one vantage point, one half hour, one resolver. Two domains came back NXDOMAIN from my machine; that is evidence they don't resolve publicly, not proof. A CDN edge can also be configured per-colo, and I didn't test that.
And I only asked about two hostnames per domain, apex and www. The fleet publishes far more subdomain URLs than that. If the apex directive has no includeSubDomains, every one of those other hostnames is a separate configuration that has to be right on its own — and I have shown that at least one of the two I did check is not.
The fix here is small: a redirect rule, two keywords on a header that already has the right number in it, and four lines nobody has written yet. The reason it is worth writing about is that all 111 sites looked fine from a browser. The padlock was there. It's the first request that wasn't.
Top comments (0)